YOLO-SR and GeoTrack for distress target detection and tracking in UAV maritime search and rescue

Reliable detection and tracking of distress targets in maritime search and rescue (SAR) remain challenging because of the extreme scale variation associated with UAV operating altitudes, where swimmers and other small targets may occupy only a few pixels. This challenge is further compounded by sea-surface reflections, wave patterns, and dynamic ocean backgrounds. This paper presents an integrated detection–tracking framework tailored to small-object maritime SAR scenarios. The proposed YOLO-SR (Small Refinement) detector extends YOLOv8-L with a lightweight high-resolution refinement branch operating on P3 feature maps to enhance the representation of small and distant targets while maintaining computational efficiency. In addition, GeoTrack extends ByteTrack through geometric filtering and motion-consistency based confidence modulation to suppress false trajectories associated with wave patterns and sea-surface reflections. Experimental evaluation on the SeaDronesSee dataset shows that YOLO-SR achieves a mAP@50 of 0.8959, mAP@50–95 of 0.5506, precision of 0.9146, and recall of 0.8724, improving upon the baseline YOLOv8-L detector. For multi-object tracking, GeoTrack achieves a MOTA of 58.4%, HOTA of 64.2%, and IDF1 of 64.2%, with improvements over the baseline ByteTrack in MOTA and IDF1. GeoTrack also achieves a tracking precision of 90.0% and reduces the proportion of Mostly Lost (ML) trajectories to 18%. The proposed framework achieves real-time inference at approximately 38 FPS on a desktop GPU and 15 FPS on an NVIDIA Jetson Nano, indicating its potential for embedded UAV-based maritime SAR applications.